Claude Opus 5.5 cuts token prices 20%. Rerun your math.

Anthropic released Claude Opus 5.5 today, and the interesting number is not the benchmark — it is the invoice. Opus 5.5 lands at $4 per million input tokens and $20 per million output, down from Opus 5's $5 and $25. Cache reads drop from $0.50 to $0.20. If you are running anything on Opus 5 in production, your cost model is stale as of this morning.

What actually happened

Anthropic announced Opus 5.5 as the first model in the Claude 5.5 family, available now on the Claude Platform, AWS, Google Cloud and Microsoft Azure under the model ID claude-opus-5-5. Anthropic says it costs 40% less to run on typical workloads than Opus 5 — that is the 20% per-token cut compounded with fewer tokens spent to finish the same job — and generates output more than 30% faster.

The benchmark deltas Anthropic published are real gaps, not rounding. Terminal-Bench 4.0 goes from 52.3% to 66.4%. FrontierCode v1.1 from 48.0% to 54.4%. OSWorld 2.0, the computer-use benchmark, from 74.0% to 81.8%. Anthropic also claims Opus 5.5 matches Fable 5.1 on coding work while costing 51% less on its HAProxy translation task.

Practical details worth reading twice: there is a faster mode billed at $8 input / $40 output, cache writes stay at $5, and five-hour usage limits went up across Pro, Max, Team and seat-based Enterprise plans. TechCrunch reports Sonnet 5.5 and Haiku 5.5 are coming in the following weeks, so the cheaper tiers have not repriced yet.

Why a 20% token price cut matters for your business

Most small companies running AI picked a model once, wired it in, and never revisited it. That was defensible when prices moved twice a year. It is not defensible now — this is the second Opus price change we have written about this quarter, and the cache-read cut is the bigger one for anyone doing retrieval.

Cache reads at $0.20 per million change the economics of stuffing context. If you have been trimming system prompts and document context to control spend, that constraint just loosened by 60%. Reference material you excluded on cost grounds may now be worth including — and the quality lift from better context usually beats the quality lift from a bigger model.

Three things to do this week:

Pull your last 30 days of token spend and split it by input, output and cache. If you cannot split it, that is the first thing to fix. You cannot act on a 60% cut to a line item you do not track.

Do not pin your integration to a dated model string and forget it. Pin deliberately, then schedule the review. A pin you chose is engineering. A pin you inherited is debt.

Re-run your eval set before you switch. Benchmarks are Anthropic's workload, not yours. A model that is better on Terminal-Bench can still be worse on your support triage prompts. Twenty test cases and an afternoon settles it.

We build AI systems where the model is a config value, not a rewrite — because vendors reprice on their schedule, not yours.

Sources: Anthropic, TechCrunch.